Many single-cell experiments can measure only a few dozen genes. Imaging-based spatial methods such as MERFISH are limited by cost and by how many probes physically fit on a sample, so panels often stay under 50 genes. Choosing that panel is a subset-selection problem over roughly 20,000 genes. Current tools take one of two routes. Some rank genes one at a time. Others train a neural network to pick genes jointly. Neither checks whether a chosen gene adds anything that the rest of the panel does not already carry, and neither is built to find pairs of genes that are useless alone but decisive together.
This note proposes a panel designer built on one statistical question: is this gene independent of cell type once we know the genes already in the panel? If the answer is yes, the gene is redundant and every candidate panel containing it can be skipped. If the answer is no, and the gene only becomes informative jointly with another gene, the pair can be evaluated as a unit. The difficulty is that single-cell data are full of false zeros ("dropouts", genes that are expressed but were not captured by sequencing). Dropouts corrupt ordinary independence tests. We rely on the one result in the file showing how to keep such tests valid under dropouts, and we study whether it survives the jump from gene-network inference to panel design.
Core Claim
The claim has two parts. First, low-order conditional independence (CI) tests can be run on raw counts with test-wise deletion: for each test, drop the cells in which any conditioning gene reads zero. Such tests can reliably flag redundant genes relative to a growing panel. Second, using these flags to shrink the candidate space makes it affordable to evaluate irreducible gene pairs as acquisition units. An irreducible subset is one in which no member is redundant given the others. Evaluating these pairs recovers cooperating markers that greedy one-gene-at-a-time selection misses. We expect the gain to appear on hard problems, meaning subtle T-cell subtypes, rare populations, and transfer across datasets. We do not expect it on easy atlases, where almost any panel works.
What Exists and Where This Sits
The pruning logic comes from the Value of Information Lattice, which organizes feature subsets into a lattice of irreducible subsets and prunes reducible ones using d-separation, a graph rule for reading off conditional independence. On the Thyroid dataset it shrank more than a million subsets to 28,806 irreducible nodes. Its look-ahead policy beat greedy acquisition, which suffers from "reluctant greed", the refusal to pay the up-front cost of multi-part tests (Bilgic and Getoor (2011); Bilgic and Getoor (2007)). d-SAGE applies the same idea to Shapley-style feature importance. It skips coalitions in which a feature is d-separated from the target, saving 38% of evaluations on Drug Consumption without bias (Luther et al. (2023)). Bayesian-network classifiers have been shown to tolerate removal of over 70% of features (Zhang et al. (2019)), which motivates the expectation that gene panels contain large redundancy. All three papers use small tabular datasets with a learned global graph. We deliberately do not learn a global graph over thousands of genes. Instead we run direct local CI tests relative to the current panel. This avoids the graph-learning bottleneck the file flags as the main failure mode of CI pruning.
The dropout handling comes from Dai et al. (2024). They model dropouts as self-masking, meaning a gene's true level sets its chance of reading zero. Under that model the marginal distributions cannot be identified, but CI statements stay valid if one conditions only on cells where the conditioning genes are nonzero. They also show that imputation with MAGIC or mixedCCA creates spurious dependencies. Their work supplies our test but also constrains us: the effective cell count collapses once more than two or three genes are conditioned on. They use the test only for pairwise regulatory edges. We use it for gene-versus-label redundancy with a capped conditioning order, and we measure how often this conditioning order cap causes wrong prunes.
The marker-selection competitors are MarkerMap and BigSur. MarkerMap relaxes discrete gene choice into a differentiable sampling layer (Gumbel-Softmax with temperature annealing). It picks fewer than 50 genes for classification or reconstruction (Gregory et al. (2024); Sarwar et al. (2022)). It gives no guarantee against redundancy and depends on how the temperature is scheduled. BigSur is a univariate filter. It separates biological from Poisson technical noise with a corrected Fano factor, keeps 100–200 genes, and processes 100,000 cells in about 90 seconds (Dollinger et al. (2025)). Its authors also document the "easy dataset trap", where random genes separate easy datasets and low-information genes set a noise floor on hard ones. We use that finding to choose where to evaluate. The review Feature selection revisited in the single-cell era notes that wrapper methods, which search subsets by retraining a model, stall beyond a few hundred genes. It does not propose CI pruning as a remedy.
Research Program
The method has three stages.
Stage one builds a cheap candidate pool. BigSur keeps the top 500 genes by corrected Fano factor, so every method starts from the same pool.
Stage two runs the dropout-aware redundancy test. For a candidate gene $g$ and the current panel $P$, we test $g \perp Y \mid S$. Here $Y$ is the cell-type label and $S \subseteq P$ holds at most two panel genes, chosen as the ones most correlated with $g$ among cells where both are nonzero. Only cells with all genes in $S$ nonzero are kept. The zeros of $g$ itself are kept, as in Dai et al. The test statistic is a conditional mutual information estimate on binned counts, with the p-value computed by permutation within strata of $S$. If fewer than a minimum number of cells survive in any cell type, the test is not run and $g$ is kept. We set that minimum from the observation that reference classifiers plateau near 500 cells per type (Abdelaal et al. (2019)), and we sweep lower values. This rule makes the method fail safe: when the data cannot support a prune, nothing is pruned.
Stage three is a budgeted look-ahead. At each step the candidate units $U$ are single genes plus irreducible pairs. A pair is irreducible when neither gene passes the redundancy test given the other, and the pair jointly raises held-out accuracy by more than the better single gene does. The step picks
$$U^\ast = \arg\max_{U \,:\, |P|+|U| \le k} \frac{\Delta(U \mid P)}{|U|},$$
where $\Delta(U \mid P)$ is the gain in cross-validated balanced accuracy of a fixed downstream classifier when $U$ is added to panel $P$, and $k$ is the probe budget. A unit is skipped without training anything if all its genes test as independent of $Y$ given $P$, which mirrors d-SAGE's skipping. The per-gene normalization is the budgeted analogue of VOILA's cost-benefit trade-off. It lets a pair win when its combined value justifies two probes, which is exactly the case greedy selection is reluctant to pay for.
A key assumption is that the cell-type label is not itself affected by dropout. The Dai et al. guarantee is stated for variables under the self-masking model. We check empirically whether it carries over to a label variable, using the calibration test below.
Baselines
The naive reference points are a random panel of $k$ genes from the pool and the top-$k$ BigSur genes (Dollinger et al. (2025)). The random panel exposes the easy-dataset trap. BigSur is the strongest univariate filter in the file and the source of our pool, so any gain over it must come from the multivariate stages. Both use the pool, budgets $k \in \{10, 25, 50\}$, and downstream classifier (a random forest, matching the best-performing family in Abdelaal et al. (2019)). We also report a full-pool random forest as the accuracy ceiling.
The strongest existing method is supervised MarkerMap (Gregory et al. (2024)). It is the only method in the file that selects nonlinear markers jointly in the fewer-than-50-gene regime, and it is reported to match or beat SMaSH, RankCorr and LassoNet on CITE-seq and Paul. We run it from its released settings on the same 500-gene pool. It gets the same number of hyperparameter configurations for its temperature schedule as our method gets for its CI significance level and minimum cell count, and its selected genes are scored by the same random forest. We found no reported panel-accuracy numbers for MarkerMap on Zeisel, Paul, CITE-seq or SSv4 on Lacuna; the only quantitative figure is runtime (about five minutes for 40,000 cells on a standard GPU). We will therefore rerun it ourselves. Beating it means higher balanced accuracy or rare-type recall at $k = 10$ or $25$ on at least two of the three hard settings described below, with non-overlapping confidence intervals across 10 splits.
The strongest simple competitor is plain greedy forward selection with the same random forest and no CI tests. The ablations that isolate the proposal's ingredients are these. Greedy with pair units but no pruning is run on a 150-gene pool, where evaluating all pairs exhaustively is affordable, and serves as an upper reference for what pruning might lose. The full method with naive CI tests on raw counts (no deletion) and with MAGIC-imputed tests measures the value of test-wise deletion. The full method without pair units measures the value of the look-ahead. Qi et al.'s whole-transcriptome MLP (81.4%) and Transformer (84.9%) accuracies on CELLxGENE (Qi et al. (2025)) come from a different task. We cite them only as context for what full-gene models reach, not as a panel baseline.
Why not just…?
Why not just rank genes with BigSur or highly variable gene filters? On easy datasets this probably is enough, and our Zeisel results may show no gap. The BigSur authors' own noise-floor finding says univariate ranking breaks down on subtle subtypes (Dollinger et al. (2025)). A filter cannot, by construction, notice that its 10th gene duplicates its 3rd. The hard-setting experiments test this directly.
Why not just use MarkerMap? It is the real competitor for accuracy. Our distinct predictions are narrower. We expect more stable panels across random seeds, because selection is discrete rather than temperature-dependent. We expect lower within-panel redundancy, measured as mean conditional mutual information between panel genes given the label. We expect better recovery of planted cooperating pairs. If MarkerMap matches us on all three, the proposal's contribution shrinks to interpretability, and we will report that.
Why not just do greedy forward selection, or greedy with pair look-ahead? This is the most serious threat. Greedy selection with a strong classifier may already find most useful genes, and exhaustive pair look-ahead over 500 genes is only about 125,000 pairs per step. The design addresses this directly. The unpruned pair-look-ahead reference on 150 genes shows whether pruning loses accuracy. The count of classifier trainings shows what pruning saves. If pruning neither saves much nor preserves accuracy, CI pruning is not worth it here.
Why not just impute dropouts and run standard tabular selection? Dai et al. (2024) show that imputation invents dependencies that produce false structure. Our MAGIC-imputed ablation tests whether the same distortion causes wrong prunes and wrong pairs in panel design, a question their edge-level results do not answer.
Proposed Experiments
The datasets come from the file. Zeisel serves as an easy control. Paul bone marrow and CITE-seq cord blood match MarkerMap's settings. The Abdelaal PBMC data provide the hard, overlapping T-cell subtypes, and the CD8+ T-cell subset from the BigSur study adds another hard case. The Baron, Muraro and Xin pancreas data support a transfer test: panels are selected on Baron and scored on Muraro and Xin, which checks whether non-redundant panels travel better across labs. The metrics are balanced accuracy, macro F1, recall on the rarest cell types, panel stability (Jaccard overlap across seeds), within-panel redundancy, number of classifier trainings, and wall-clock time.
The first decisive check is a planted-synergy simulation using the BoolODE and BEELINE simulators from Dai et al. We define one subtype by an XOR-like rule on two genes. We add near-copies of informative genes and apply self-masking dropouts at increasing rates. Success means the full method recovers the planted pair when greedy selection and BigSur do not, and that deletion-based tests keep the rate of wrong prunes near the nominal significance level while naive and imputed tests exceed it.
The second decisive check is calibration of pruning on real data. For every gene pruned at every step, we add it back and measure the held-out accuracy gain. If pruning is sound, this distribution should be centered at zero. We report how it degrades as the minimum cell count falls and the conditioning order rises from one to three. This maps where test-wise deletion stops being usable for panel design, which is the open power question the file leaves unresolved.
The final comparison is at $k = 10$ and $25$ on PBMC, CD8+ T cells and pancreas transfer. It pits the full method against MarkerMap and greedy selection under equal tuning budgets and counts classifier trainings saved relative to the unpruned pair reference.
References
- Value of Information Lattice: Exploiting Probabilistic Independence for Effective Feature Subset Acquisition
- VOILA: Efficient Feature-value Acquisition for Classification
- Efficient SAGE Estimation via Causal Structure Learning
- Structure Learning of Bayesian Network Based on Adaptive Thresholding
- Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View
- MarkerMap: nonlinear marker selection for single-cell studies
- MarkerMap: nonlinear marker selection for single-cell studies
- Statistically principled feature selection for single cell transcriptomics
- Feature selection revisited in the single-cell era
- Evaluation of single-cell classifiers for single-cell RNA sequencing data sets
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics